Since #28140 when the global checkpoint is advanced, we try to move the
safe commit forward, and clean up old index commits if possible. However,
we forget to trim unreferenced translog.
This change makes sure that we prune both old translog and index commits
when the safe commit advanced.
Relates #28140Closes#32089
Subclasses of `EsIntegTestCase` run multiple Elasticsearch nodes in the
same JVM and when we log we look at the name of the thread to figure out
the node name. This makes sure that all calls to `daemonThreadFactory`
include the node name.
Closes#32574
I'd like to follow this up with more drastic changes that make it
impossible to do this incorrectly but that change is much larger than
this and I'd like to get these log lines fixed up sooner rather than
later.
All Translog inner closes should happen after tragedy exception is set (#32674)
We faced with the nasty race condition. See #32526
InternalEngine.failOnTragic method has thrown AssertionError.
If you carefully look at if branches in this method, you will spot that its only possible, if either Lucene IndexWriterhas closed from inside or Translog, has closed from inside, but tragedy exception is not set.
For now, let us concentrate on the Translog class.
We found out that there are two methods in Translog - namely rollGeneration and trimOperations that are closing Translog in case of Exception without tragedy exception being set.
This commit fixes these 2 methods. To fix it, we pull tragedyException from TranslogWriter up-to Translog class, because in these 2 methods IndexWriter could be innocent, but still Translog needs to be closed. Also, tragedyException is wrapped with TragicExceptionHolder to reuse CAS/addSuppresed functionality in Translog and TranslogWriter.
Also to protect us in the future and make sure close method is never called from inside Translog special assertion examining stack trace is added. Since we're still targeting Java 8 for runtime - no StackWalker API is used in the implementation.
In the stack-trace checking method, we're considering inner caller not only Translog methods but Translog child classes methods as well. It does mean that Translog is meant for extending it, but it's needed to be able to test this method.
Closes#32526
This is related to #32517. This commit passes the DiscoveryNode to the
initiateChannel method for different Transport implementation. This
will allow additional attributes (besides just the socket address) to be
used when opening channels.
Randomized test conditions that cause some shards to have no docs on them
failed due to test asserts that relied on a lazy initialization side effect
from the map script. After this fix:
- Test cases with the relevant init script are protected
- Test cases with the relevant combine or reduce scripts were already
protected, because the combine and reduce scripts safely handle this case.
This is a followup to #31886. After that commit the
TransportConnectionListener had to be propogated to both the
Transport and the ConnectionManager. This commit moves that listener
to completely live in the ConnectionManager. The request and response
related methods are moved to a TransportMessageListener. That listener
continues to live in the Transport class.
* Lazy resolve DNS (i.e. `String` to `DiscoveryNode`) to not run into indefinitely caching lookup issues (provided the JVM dns cache is configured correctly as explained in https://www.elastic.co/guide/en/elasticsearch/reference/6.3/networkaddress-cache-ttl.html)
* Changed `InetAddress` type to `String` for that higher up the stack
* Passed down `Supplier<DiscoveryNode>` instead of outright `DiscoveryNode` from `RemoteClusterAware#buildRemoteClustersSeeds` on to lazy resolve DNS when the `DiscoveryNode` is actually used (could've also passed down the value of `clusterName = REMOTE_CLUSTERS_SEEDS.getNamespace(concreteSetting)` together with the `List<String>` of hosts, but this route seemed to introduce less duplication and resulted in a significantly smaller changeset).
* Closes#28858
Currently, if geo context is represented by something other than
geo_point or an object with lat and lon fields, the parsing of it
as a geo context can result in ignoring the context altogether,
returning confusing errors such as number_format_exception or trying
to parse the number specifying as long-encoded hash code. It would also
fail if the geo_point was stored.
This commit makes the mapping parsing more strict and will fail during
mapping update or index creation if the geo context doesn't point to
a geo_point field.
Supersedes #32412Closes#32202
* Scripted metric aggregations: add deprecation warning and system property to control legacy params
Scripted metric aggregation params._agg/_aggs are replaced by state/states context variables. By default the old params are still present, and a deprecation warning is emitted when Scripted Metric Aggregations are used. A new system property can be used to disable the legacy params. This functionality will be removed in a future revision.
* Fix minor style issue and docs test failure
* Disable deprecated params._agg/_aggs in tests and revise tests to use state/states instead
* Add integration test covering deprecated scripted metrics aggs params._agg/_aggs access
* Disable deprecated params._agg/_aggs in docs integration tests and revise stored scripts to use state/states instead
* Revert unnecessary migrations doc change
A relevant note should be added in the changes destined for 7.0; this PR is going to be backported to 6.x.
* Replace deprecated _agg param bwc integration test with a couple of unit tests
* Fix compatibility test after merge
* Rename backwards compatibility system property per code review feedback
* Tweak deprecation warning text per review feedback
This fix prevernts trying to parse unknown timezone ids by converting
the joda time zone via java.util.TimeZone to a java time based ZoneId.
Closes#32927
testDocStats test is flaky and sometimes it's failing on jenkins and
failure is not reproducible locally. The reason for this failure is in
timing. If the number of deleted documents is greater than 33% of inserted
documents, Lucene will schedule segments to merge if TieredMergePolicy is
used (it's not the case for LogMergePolicy, but ES is only using
TieredMergePolicy). If this merge is performed before stats are
retrieved - we will get 0 for "deleted" counter.
So basically this counter could be either 0 or numOfDeletedDocs at this point,
but this is the too loose assertion and we decided to remove it at all.
Closes#32766
This commit disables the automatic `refresh_interval` in order to ensure
that index readers cannot differ between the normal and scroll search.
This issue is related to the 7.5 Lucene upgrade which contains a change that
makes single segment merge more likely to occur (max deletes percentage).
Closes#32682
We do not support passphrases on the secure settings storage (the
keystore). Yet, we added support for this in the API layer. This commit
removes this support so that we are not limited in our future options,
or have to make a breaking change.
This change cleans up some methods in the CharArrays class from x-pack, which
includes the unification of char[] to utf8 and utf8 to char[] conversions that
intentionally do not use strings. There was previously an implementation in
x-pack and in the reloading of secure settings. The method from the reloading
of secure settings was adopted as it handled more scenarios related to the
backing byte and char buffers that were used to perform the conversions. The
cleaned up class is moved into libs/core to allow it to be used by requests
that will be migrated to the high level rest client.
Relates #32332
This commit fixes a global checkpoint listeners test wherein we were
expecting an executor to have been used even if there were no
listeners. This is silliness, so this commit adjusts the assertion to
verify that the executor never fires if there are no listeners, and
fires exactly once if there is one or more listeners.
This commit introduces the ability for global checkpoint listeners to be
registered at the shard level. These listeners are notified when the
global checkpoint is updated, and also when the shard closes. To
encapsulate these listeners, we introduce a shard-level component that
handles synchronization of notification and modifications to the
collection of listeners.
This is related to #31835. It moves the default connection profile into
the ConnectionManager class. The will allow us to have different
connection managers with different profiles.
This removes custom Response classes that extend `AcknowledgedResponse` and do nothing, these classes are not needed and we can directly use the non-abstract super-class instead.
While this appears to be a large PR, no code has actually changed, only class names have been changed and entire classes removed.
This commit adds a java time version of the existing rounding classes, which features the same test suite and a small test class to check if serialization works as expected.
Significance score doubles were being parsed as long. Existing tests did not catch this because SignificantLongTermsTests and SignificantStringTermsTests did not set the score. Fixed these and also added integration test.
Thanks for the report/fix, Blakko
Closes#32770
* INGEST: Create Index Before Pipeline Execute
* Ensures that indices are created before the default pipeline setting is read to correcly handle the case of an index template containing a default pipeline (without the fix the first document does not get the pipeline applied as explained in #32758)
* closes#32758
#31821 introduced an unreleased bug where NOOP updates were incorrectly mutating the bulk
shard request, inserting null item to be replicated, which would result in NullPointerExceptions when
serializing the request to be shipped to the replicas.
Closes#32808
This is related to #31835. This commit adds a connection manager that
manages client connections to other nodes. This means that the
TcpTransport no longer maintains a map of nodes that it is connected
to.
With the move to java time, the default formatter used by toString on
ZonedDateTime uses optional components for least significant portions of
the date. This commit changes the cat indices api to use a strict date
time format, which will always output milliseconds, even if they are
zero.
closes#32466
Currently AbstractBuilderTestCase generates certain random values in its
`beforeTest()` method annotated with @Before only the first time that a test
method in the suite is run while initializing the serviceHolder that we use for
the rest of the test. This changes the values of subsequent random values
and has the effect that when running single methods from a test suite with
"-Dtests.method=*", the random values it sees are different from when the same
test method is run as part of the whole test suite. This makes it hard to use
the reproduction lines logged on failure.
This change runs the inialization of the serviceHolder and the randomization
connected to it using the test runners master seed, so reproduction by running
just one method is possible again.
Closes#32400
Processing bulk request goes item by item. Sometimes during processing, we need to stop execution and wait for a new mapping update to be processed by the node. This is currently achieved by throwing a `RetryOnPrimaryException`, which is caught higher up. When the exception is caught, we wait for the next cluster state to arrive and process the request again. Sadly this is a problem because all operations that were already done until the mapping change was required are applied again and get new sequence numbers. This in turn means that the previously issued sequence numbers are never replicated to the replicas. That causes the local checkpoint of those shards to be stuck and with it all the seq# based infrastructure.
This commit refactors how we deal with retries with the goal of removing `RetryOnPrimaryException` and `RetryOnReplicaException` (not done yet). It achieves so by introducing a class `BulkPrimaryExecutionContext` that is used the capture the execution state and allows continuing from where the execution stopped. The class also formalizes the steps each item has to go through:
1) A translation phase for updates
2) Execution phase (always index/delete)
3) Waiting for a mapping update to come in, if needed
4) Requires a retry (for updates and cases where the mapping are still not available after the put mapping call returns)
5) A finalization phase which allows updates to the index/delete result to an update result.
This adds a java time based date math parser class in order, which will replace the joda date based one in the future. For now the class also returns the date in milliseconds since the epoch.
Currently if a document cannot be indexed because it violates the defined
mapping for the index, a MapperException is thrown. In some cases it is
useful to expose the expected field type in the exception itself,
so that the user can react based on the error message. This change adds
the expected data type to the MapperException.
Closes#31502
Remove a few of the logger constructors that aren't widely used or
aren't used at all and deprecate a few more logger constructors in favor
of log4j2's `LogManager`.
A bug in the test suite prevented to properly check that all date
formatters printed the date the same way like joda time does.
This fixes the test and thus also a fair share of formats, that
now use the strict parser for printing.
We previously discussed moving the classes extending `AcknowledgedResponse` to
simply use `AcknowledgedResponse`, making the class non-abstract.
This moves the first class to do this, removing `WritePipelineResponse` in the
process.
If we like the way this looks, I will switch the remaining classes over to using
`AcknowledgedResponse`.
When Circuit Breaker has tripped, certain diagnostic requests like
"_cluster/health" succeed where as request to / fails with
503 Service Unavailable. This behavior is observed because of this
commit f32b700 where certain API paths are whitelisted from
Circuit Breaking exception, but / is not whitelisted.
Added / to circuit breaker whitelist so that it can be used for
diagnostic purposes
* Fixes suggestion generics
This solves a compile problem in Eclipse where Eclipse could not
resolve the generics for the options field in `PhraseSuggestion.Entry`.
But I think this is also a good change in general because
`PhraseSuggestion.Entry` is now declaring the specific `Option`
implementation it requires rather than `Suggest.Entry.Option` which is
more general and could lead to weird bugs. `CompletionSuggestion.Entry`
and `TermSuggestion.Entry` already declare the more specific class they
use so I think this was an oversight in `PhaseSuggestion.Entry`
* iter
`ShardOperationFailedException` and corresponding implementors seem to suggest that the cause may be null, case that is also handled in a few places. Yet, it does not seem to be possible in practice for the cause to be null, hence we can clean that up and enforce the cause to be a non null value. This is best done by making `ShardOperationFailedException` an abstract class rather than an interface, which holds the basic member instance that all the subclasses have in common and can also enforce that cause, status and reason are non null.
As part of #32608 we made sure that the fully qualified index name is taken from the query shard context whenever creating a new `QueryShardException`. That change introduced a regression as instead of setting the entire `Index` object to the exception, which holds index name and index uuid, we ended up setting only the index name (including cluster alias). With this commit we make sure that the index uuid does not get lost and we try to lower the chances that a similar bug makes it in another time. That's done by making `QueryShardContext` return the fully qualified `Index` (which also holds the uuid) rather than only the fully qualified index name.
Suggestion responses were previously serialized as streamables which
made writing suggesters in plugins with custom suggestion response types
impossible. This commit makes them serialized as named writeables and
provides a facility for registering a reader for suggestion responses
when registering a suggester.
This also makes Suggestion responses abstract, requiring a suggester
implementation to provide its own types. Suggesters which do not need
anything additional to what is defined in Suggest.Suggestion should
provide a minimal subclass.
The existing plugin suggester integration tests are removed and
replaced with an equivalent implementation as an example
plugin.
It will be useful for future efforts to know if the global checkpoint
was updated. To this end, we need to expose whether or not the global
checkpoint was updated when the state of the replication tracker
updates. For this, we add to the tracker a callback that is invoked
whenever the global checkpoint is updated. For primaries this will be
invoked when the computed global checkpoint is updated based on state
changes to the tracker. For replicas this will be invoked when the local
knowledge of the global checkpoint is advanced from the primary.
The MockNioTransport (similar to the MockTcpTransport) is used for integ
tests. The MockTcpTransport has always only opened a single for all of
its work. The MockNioTransport has awlays opened the default number of
connections (13). This means that every test where two transports
connect requires 26 connections. This is more than is necessary. This
commit modifies the MockNioTransport to only require 3 connections.
Primary terms were introduced as part of the sequence-number effort (#10708) and added in ES
5.0. Subsequent work introduced the replication tracker which lets the primary own its replication
group (#25692) to coordinate recovery and replication. The replication tracker explicitly exposes
whether it is operating in primary mode or replica mode, independent of the ShardRouting object
that's associated with a shard. During a primary relocation, for example, the primary mode is
transferred between the primary relocation source and the primary relocation target. After
transferring this so-called primary context, the old primary becomes a replication target and the
new primary the replication source, reflected in the replication tracker on both nodes. With the
most recent PR in this area (#32442), we finally have a clean transition between a shard that's
operating as a primary and issuing sequence numbers and a shard that's serving as a replication
target. The transition from one state to the other is enforced through the operation-permit system,
where we block permit acquisition during such changes and perform the transition under this
operation block, ensuring that there are no operations in progress while the transition is being
performed. This finally allows us to turn the best-effort checks that were put in place to prevent
shards from being used in the wrong way (i.e. primary as replica, or replica as primary) into hard
assertions, making it easier to catch any bugs in this area.
Currently, when TranslogCorruptedException is thrown most of the times it does not contain information about the translog location on the file system. There is the translog recovery tool that accepts the translog path as an argument and users are constantly puzzled where to get the path.
This pull request adds "source" information to every TranslogCorruptedException thrown. The source could be local file, remote translog source (used for recovery), assertion (translog entry is constructed to perform some assertion) or translog constructed inside the test.
Closes#24929
This change adds a check so that when parsing the search source, script fields are
ignored when the requested search result size is 0. This helps with e.g. clients like
Kibana that sends a list of script fields that they may need for convenience, but they
don't require any hits. Before this change, user sometimes ran into confusing behaviour,
e.g. the script compilation limit to breaking although no hits were requested.
Closes#31824
* We were comparing the wrong timeout value in the `randomValueOtherThan` call here, leading to no mutation happening for a certain seed
* closes#32639
Today content type detection on an input stream works by peeking up to
twenty bytes into the stream. If the stream is headed by more whitespace
than twenty bytes, we might fail to detect the content type. We should
be ignoring this whitespace before attempting to detect the content
type. This commit does that by ignoring all leading whitespace in an
input stream before attempting to guess the content type.
The assertion in the test was not broad enough. If the timing is very unlucky, the
shard is already promoted to primary before the indexOnReplica even gets to execute.
Closes#32645
* Remove UpdateSettingsTestHelper class
By making the `settings()` method public on `UpdateSettingsRequest` (I think it
should have been in the first place) we can get rid of this class entirely. Mock
response objects are now constructed by parsing JSON without making the
constructor public.
Relates to #29823
* Remove RolloverIndexTestHelper
This removes the `RolloverIndexTestHelper` class in favor of making a couple of
getters publically accessible as well as custom building a response object using
JSON parsing.
Relates to #29823
This commit removes the hacks associated with mocking Response objects. Rather
than parse a wrapped byte array, the constructors for `IndicesAliasesResponse`
and `ResizeResponse` are made public
Relates to #29823
When some remote clusters return shard failures as part of a cross-cluster search request, the cluster alias currently gets lost. As a result, if the shard failures are all caused by the same error, and against indices belonging to different clusters, but with the same index name, only one failure gets returned as part of the search response, meaning that failures are grouped by index name, ignoring the cluster alias.
With this commit we make sure that `ShardSearchFailure` returns the cluster alias as part of the index name. Also, we set the fully qualfied index name when creating a `QueryShardException`. That way shard failures are grouped by cluster:index. Such fixes should cover at least most of the cases where either 1) the shard target is set but we don't have the index in the cause (we were previously reading it only from the cause that did not have the cluster alias) 2) the shard target is missing but if the cause is a `QueryShardException` the cluster alias does not get lost.
We also prevent NPE in case the failure cause is not set and test such scenario.
If the shard is already closed while bumping the primary term, this can result in an
AlreadyClosedException to be thrown. As we use asyncBlockOperations, the exception
will be thrown on a thread from the generic thread pool and end up in the uncaught
exception handler, failing our tests.
Relates to #32442
Some classes use internal date formatters, which now can be moved over
to java time using the DateFormatters class.
The same applies for a few test cases.
This unmutes the testFanOutAndCollect()` method and add a check to make
sure we aren't accidentally running something twice causing a search
phase to still be running after we have counted down the latch
Relates to #29242
We've recently seen a number of test failures that tripped an assertion in IndexShard (see issues
linked below), leading to the discovery of a race between resetting a replica when it learns about a
higher term and when the same replica is promoted to primary. This commit fixes the race by
distinguishing between a cluster state primary term (called pendingPrimaryTerm) and a shard-level
operation term. The former is set during the cluster state update or when a replica learns about a
new primary. The latter is only incremented under the operation block, which can happen in a
delayed fashion. It also solves the issue where a replica that's still adjusting to the new term
receives a cluster state update that promotes it to primary, which can happen in the situation of
multiple nodes being shut down in short succession. In that case, the cluster state update thread
would call `asyncBlockOperations` in `updateShardState`, which in turn would throw an exception
as blocking permits is not allowed while an ongoing block is in place, subsequently failing the shard.
This commit therefore extends the IndexShardOperationPermits to allow it to queue multiple blocks
(which will all take precedence over operations acquiring permits). Finally, it also moves the primary
activation of the replication tracker under the operation block, so that the actual transition to
primary only happens under the operation block.
Relates to #32431, #32304 and #32118
* Make cluster stats response contain cluster UUID
* Updating constructor usage in Monitoring tests
* Adding cluster_uuid field to Cluster Stats API reference doc
* Adding rest api spec test for expecting cluster_uuid in cluster stats response
* Adding missing newline
* Indenting do section properly
* Missed a spot!
* Fixing the test cluster ID
Unmuting the test and adding some more debug output. Was not able to
reproduce the prior failure, but it seems possible that the
failure (mismatched counts) could be caused by partial search results
during the test.
The assertions check for shard failures first, because if one of the
two searches is partial the rest of the test will fail.
Next, instead of just checking respective hit counts, we emit the
difference in hits to help identify what went wrong.
Closes#32492
Since LUCENE-8263, testSeqNoAndCheckpoints might trigger merges because
of the updates and deletes in the test. Our merge scheduler will trigger
a flush if there is no pending merge. Those extra flushes will change
the last committed segmentInfos in the engine and fail the test.
This commit uses LogMergePolicy for the engine in the test to avoid
merges.
Closes#32430
This commit adds a boolean system property, `es.scripting.use_java_time`,
which controls the concrete return type used by doc values within
scripts. The return type of accessing doc values for a date field is
changed to Object, essentially duck typing the type to allow
co-existence during the transition from joda time to java time.
This removes a constructor from `AbstractComponent` and
`AbstractLifecycleComponent` that we weren't using and it switches the
logger creation away from one of the `Settings` flavored methods which
are no longer needed.
First, some background: we have 15 different methods to get a logger in
Elasticsearch but they can be broken down into three broad categories
based on what information is provided when building the logger.
Just a class like:
```
private static final Logger logger = ESLoggerFactory.getLogger(ActionModule.class);
```
or:
```
protected final Logger logger = Loggers.getLogger(getClass());
```
The class and settings:
```
this.logger = Loggers.getLogger(getClass(), settings);
```
Or more information like:
```
Loggers.getLogger("index.store.deletes", settings, shardId)
```
The goal of the "class and settings" variant is to attach the node name
to the logger. Because we don't always have the settings available, we
often use the "just a class" variant and get loggers without node names
attached. There isn't any real consistency here. Some loggers get the
node name because it is convenient and some do not.
This change makes the node name available to all loggers all the time.
Almost. There are some caveats are testing that I'll get to. But in
*production* code the node name is node available to all loggers. This
means we can stop using the "class and settings" variants to fetch
loggers which was the real goal here, but a pleasant side effect is that
the ndoe name is now consitent on every log line and optional by editing
the logging pattern. This is all powered by setting the node name
statically on a logging formatter very early in initialization.
Now to tests: tests can't set the node name statically because
subclasses of `ESIntegTestCase` run many nodes in the same jvm, even in
the same class loader. Also, lots of tests don't run with a real node so
they don't *have* a node name at all. To support multiple nodes in the
same JVM tests suss out the node name from the thread name which works
surprisingly well and easy to test in a nice way. For those threads
that are not part of an `ESIntegTestCase` node we stick whatever useful
information we can get form the thread name in the place of the node
name. This allows us to keep the logger format consistent.
When using cross-cluster search through the high-level REST client, the cluster alias from each search hit was not parsed correctly. It would be part of the index field initially, but overridden just a few lines later once setting the shard target (in case we have enough info to build it from the response). In any case, getClusterAlias returns `null` which is a bug.
With this change we rather parse back clusterAliases from the index name, set its corresponding field and properly handle the two possible cases depending on whether we can or cannot build the shard target object.
Nhat Nguyen
Nik Everett
Andrey Ershov
Tim Brooks
Colin Goodheart-Smithe
Jonathan Little
Alpar Torok
Armin Braun
Tal Levy
Igor Motov
Alexander Reelsen
Paul Sanwald
JB Nizet
Julie Tibshirani
Jim Ferenczi
Jason Tedor
Adrien Grand
Jay Modi
Lee Hinman
Andy Bristol
markharwood
Yannick Welsch
Ryan Ernst
Christoph Büscher
Boaz Leskes
lipsill
Nicholas Knize
Suresh N S
Luca Cavanna
Parth Verma
Lee Hinman
Shaunak Kashyap
Zachary Tong
Sohaib Iftikhar
Daniel Mitterdorfer